Models Whose Weights Are Published
Open weights describes an LLM whose weights, the core of the model, are published, so you can place it on your own computer or server and run it. Weights are the large set of numbers fixed by training, and with them you can run the model. Models whose weights are not published are used through their developer's app or API. On the pages of this site, the "Open weights only" filter narrows the list to models with published weights.
The advantages of running a model yourself are that you do not have to send documents to an outside service and that you pay nothing per use. It suits having a model read internal documents and answer, or adjusting a model yourself for Japanese. The larger the model, however, the more machine power and electricity it takes to run.
Not the Same as Open Source
Open weights does not mean the same as open source. The OSI, which stewards the definition of open source, has set out a definition for AI (version 1.0). It requires publishing not only the weights but also detailed information about the data used for training and the complete code for training and running the system. Under this definition, a model with only its weights published does not count as open source.
Read the License Before Use
Even when the weights are published, what you may do with them differs from license to license. The Apache 2.0 license allows copying, modifying, and redistributing at no charge, and requires a copy of the license and notices of changes when you redistribute. On the other hand, Meta's Llama 3.1 license was released on July 23, 2024. It requires companies whose monthly users exceeded 700 million as of the release date to obtain a separate license from Meta. It also requires prominently displaying "Built with Llama" when redistributing.
Before running a model yourself, read the license on the model's download page and check whether commercial use is allowed and whether there are display requirements.